File size: 4,965 Bytes
8e6b0e0 906ccca 8e6b0e0 906ccca 8e6b0e0 906ccca 8e6b0e0 e4e8ce5 8e6b0e0 e4e8ce5 8e6b0e0 e4e8ce5 906ccca 8e6b0e0 906ccca e4e8ce5 906ccca e4e8ce5 8e6b0e0 906ccca 8e6b0e0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 | {
"schema_version": 1,
"runtime_image": "sha256:7447e3f5a1f1bd9f83cb60c3d4d22cd7fa3d44195b31a53982c63b988e660793",
"runtime_image_tag": "lottolabs/gemma4-12b-tt-p150:p3d-pc-gate-s1",
"gemma4_tree": {
"path": "runtime/gemma4/",
"installed_at": "/home/container_app_user/tt-metal/models/demos/gemma4/",
"git_branch": "sampling (src-g2)",
"git_commit": "d6f920246c37a9b6eeed74b6bd853da683e4ff7b",
"note": "Copied from the runtime image; equal to the git tree at that commit (diff -r); .git and __pycache__ excluded."
},
"tools": {
"path": "runtime/tools/",
"installed_at": "/home/container_app_user/gemma4-tools/"
},
"serving_overlay": {
"path": "runtime/serving-overlay/",
"installed_at": "/",
"content": "vLLM core files (speculative method gemma4_assistant, multi-token stop handling) and the complete vllm_tt_plugin package as installed in the image: exact prefix resume (prefix_resume.py, scheduler.py), on-device sampling and speculative-sampling plumbing (model_runner.py, platform.py, worker.py, model_input.py)",
"vllm_tt_plugin_git_commit": "d6f920246c37a9b6eeed74b6bd853da683e4ff7b",
"note": "Copied from the runtime image; the plugin package equals the src-g2 git tree at that commit."
},
"ttnn_overlay": {
"path": "runtime/ttnn-overlay/",
"installed_at": "/home/container_app_user/tt-metal/ttnn/",
"git_branch": "dn",
"git_commit": "efc465f4b1a44025ae4a33e4dbc7d371b3f9266c",
"content": "1D matmul program factory with in0-multicast chunking and the dual-NoC in1 reader (TT_MM1D_IN1_* column masks), in0/in1 dataflow reader kernels; TTNN host libraries rebuilt in layer ttnn-dn1",
"note": "Byte-identical to the three files in the runtime image. Layer p2b built an earlier revision of the factory file (overlay master 5200a780d6feabdc387f7c106f25f43f4a85a0a8) that ttnn-dn1 overwrites."
},
"runtime_env": {
"GEMMA4_VERIFY_SDPA": "batched",
"GEMMA4_FUSE_GELU_MUL": "1",
"GEMMA4_DECODE_KV_ROWS": "1",
"GEMMA4_DECODE_MM_DN": "1",
"binding": "Set in the image (layer fast2) and recorded as runtime_env in both proofs; the loader rejects any other numerics environment."
},
"build_chain": [
{
"layer": "base",
"image": "ghcr.io/tenstorrent/tt-inference-server/vllm-tt-metal-src-release-ubuntu-22.04-amd64:0.20.0-de59f8a-03fa3af",
"image_id": "sha256:710de20012a5417ec632e9d1db33b45f81fa9d9961da882a94868d15c5f99735"
},
{
"layer": "p2b",
"dockerfile": "runtime/Dockerfile.p2b",
"image_id": "sha256:9fc0b344293da678ef107d570753fea90205968a9cced44ec81831f10163cb3b"
},
{
"layer": "p2c",
"dockerfile": "runtime/Dockerfile.p2c",
"image_id": "sha256:b91c7e92ec16c507c916229ba7d675093379fe18ab05c18732eedcd3b91156f2"
},
{
"layer": "ttnn-dn1",
"dockerfile": "runtime/Dockerfile.ttnn-dn1",
"build_context": "runtime/ttnn-overlay/ as ttnn/",
"image_id": "sha256:2807668702d59ed5cfd914e76df10ab1f6276d3fce3e7e1d174040156ec5d8d0"
},
{
"layer": "fast2",
"dockerfile": "runtime/Dockerfile.fast2",
"image_id": "sha256:1c23787a9da68bdbb289a190deb86647061c36c7e5fec709be8e358e269ae4b7"
},
{
"layer": "p3d",
"dockerfile": "runtime/Dockerfile.p3d",
"image_id": "sha256:331e79b24fa2575c436158be994c6bc0e184d776d15305e82538d2e4d646f544"
},
{
"layer": "p3d-pc",
"dockerfile": "runtime/Dockerfile.p3d-pc",
"image_id": "sha256:f3b797e6bf715e30ec116ad25098352c6754aa373617459dec5e34b2ca01c2b8",
"content": "exact prefix resume (gemma4 tree + vllm_tt_plugin prefix_resume.py / scheduler.py / platform.py / model_runner.py)"
},
{
"layer": "p3d-pc-gate",
"dockerfile": "runtime/Dockerfile.p3d-pc-gate",
"image_id": "sha256:b8e8d2e86c6b10440d2127697216f567380c2e185edcafe90c1e4b6e8db57327",
"content": "confidence-gated draft length (GEMMA4_SPEC_GATE default 4096:0.85,49152:0.95)"
},
{
"layer": "p3d-pc-gate-s1",
"dockerfile": "runtime/Dockerfile.p3d-pc-gate-s1",
"image_id": "sha256:7447e3f5a1f1bd9f83cb60c3d4d22cd7fa3d44195b31a53982c63b988e660793",
"content": "on-device exact sampling, lossless speculative sampling, drafter-KV-row fix; complete gemma4 tree and vllm_tt_plugin package (git d6f9202)"
}
],
"build_script": "runtime/build.sh",
"rebuild_note": "Provenance only: intermediate layers copied earlier gemma4 / plugin trees that the p3d-pc-gate-s1 layer overwrites with complete copies; the Dockerfiles reference local intermediate tags. Serve the checksum-pinned runtime-image.tar.gz.",
"previous_runtime": {
"tag": "lottolabs/gemma4-12b-tt-p150:p3d",
"image_id": "sha256:331e79b24fa2575c436158be994c6bc0e184d776d15305e82538d2e4d646f544",
"gemma4_git_commit": "061e48ddf6a0eab5dac1f958731fcb70768231d1"
},
"licenses": {
"tt-metal": "runtime/licenses/tt-metal/",
"vllm": "runtime/licenses/vllm/"
},
"sampling_source_patches": [
"evidence/p3d-pc-gate-s1/sampling/source/gemma-sampling-s1.patch",
"evidence/p3d-pc-gate-s1/sampling/source/gemma-sampling-s2.patch"
]
}
|